video-generation

name	video-generation
description	Video generation. Animates a source image into a short video clip, or generates video from text. Use for any output that requires motion — bringing images to life, camera moves, atmospheric animation, or full scene videos.
metadata	{"type":"primitive"}

When to use

Use i2v when:

• The user wants to "animate" or "bring to life" an existing canvas image.
• A plan node turns a keyframe into a clip (e.g. scene image → scene video).
• A t2i node in the same plan produces the source frame (wire with a depends_on edge).

Always prefer i2v over t2v when a source image is available — it will be visually consistent with the reference.

---

Prompt structure

The source image fixes the subject's appearance. The prompt describes only what changes: motion, camera, light shifts, atmosphere. Never redescribe the subject.

[ACTION]

One verb. One motion arc. That is the entire action budget.

• Bad: "She walks over, turns around, and smiles at the camera."
• Good: "She turns slowly toward the lens."
• Add 1–2 micro-motions that support the main action: "fabric shifts at the shoulder," "steam rises from the cup," "hair catches the light."
• Use force verbs — avoid weightless defaults:
  • "walks" → "pushes through"
  • "looks at" → "snaps attention toward"
  • "runs" → "charges forward"

[CAMERA]

One camera move + one rhythm word. Never combine two moves — it causes jitter.

Move options: slow push-in / static tripod / horizontal pan / gradual orbit / handheld shoulder-cam / rack focus / dolly pull-back / crash zoom / whip pan

Rhythm words: gradual, gentle, smooth, abrupt, sharp, fluid

Example: "Gradual dolly push-in." or "Static tripod shot, no movement."

[ENVIRONMENT & LIGHTING]

Name every light source by type — never by mood.

• Bad: "cinematic lighting," "golden hour," "moody atmosphere," "beautiful light."
• Good: "warm tungsten side-light from the left," "motivated lighting — practical lamp visible in frame," "neon blue rim light," "overcast north-facing window light," "sodium-vapour street lamp at 2 o'clock."

Describe atmospheric changes if they evolve over the clip: "warm golden light fades to cooler dusk tone over 6 seconds."

Add realism anchors when photorealism is required: "Fine 35mm film grain. 2% camera shake in the first 1–2 seconds."

[AUDIO]

Always specify. Silent defaults produce random results.

• Dialogue (in quotes): She says: "I've been thinking about this for weeks."
• Ambient layer: "coffee shop murmur at low volume," "room tone at -28dB," "ocean wind."
• Sound effects: "ceramic cup placed on wooden table," "fabric rustle on camera pickup," "distant thunder."
• Music: "slow minor key acoustic guitar," "no music."
• Silence: "No dialogue. No music. Room tone only."

[CONSTRAINTS]

Positive phrasing only — write what you want, not what you don't.

• "Stable picture throughout."
• "Maintain outfit continuity."
• "Sharp focus on subject. Background softly defocused."
• "Consistent facial proportions throughout."
• "No morphing. No identity drift." → rephrase as: "Subject appearance stays identical to the source image throughout."

Anti-AI detection layer when realism matters:

• "2–3% film grain overlay."
• "2% camera shake in the first 1–2 seconds."
• "Slight facial asymmetry maintained from source."

[QUALITY SUFFIX]

Append to every prompt without exception: 4K. Ultra HD. Rich details. Sharp clarity. Cinematic texture. Natural colors. Stable picture.

---

Duration

MUST be exactly 4, 6, or 8 seconds — no other values.

• 4s: subtle atmosphere, minimal motion, single beat.
• 6s: one camera move or moderate subject motion.
• 8s: multi-beat action, evolving lighting, or complex camera arc.

For sequences longer than 8 seconds, split into multiple nodes connected with concat.

---

Model hints

• gemini-omni-flash-preview: default — best for general video generation, supports stateful conversational editing (audio references/mixing are NOT supported).
• veo-3.1-lite-generate-001: best balance of quality and speed for Veo model family.
• veo-3.1-generate-001: highest quality motion and consistency; use for hero shots or final output.
• Use the canvas default model unless the user explicitly requests otherwise.

---

Common failures

• Redescribing subject appearance conflicts with the source image — describe motion only.
• Two camera moves in one prompt ("pan left while tracking right") causes jitter in every model.
• Short duration (4s) with complex multi-step action loses beats — use 6s or 8s.
• Busy source images + strong camera moves cause flickering — prefer subtle moves or static tripod.
• Omitting audio leaves the model to hallucinate sound — always specify, even if the choice is silence.
• Connecting Audio Nodes: Connecting a separate audio/music node (t2m, t2s) as a reference/dependency to a video node is not supported for any model, including gemini-omni-flash-preview.
• Stateful Video Editing (Omni): The default model 'gemini-omni-flash-preview' supports stateful editing! To edit an existing video (e.g. 'make it faster', 'change the style', 'add a character'), draw a 'depends_on' edge from the previous video node to the new video node. The engine will propagate the interaction state for seamless editing.